Current methods and systems of defect detection on wafers are usually based on comparisons of signals obtained from inspection of a number of adjacent wafer dies or fields of view, featuring a like pattern. Defects produced during wafer fabrication are assumed to be random in nature. Therefore, defect detection is based on a statistical approach, whereby the probability that an identical random defect will be present at the same location within adjacent wafer dies is very low. Hence, defect detection is commonly based on identifying irregularities through the use of the well known method of die-to-die comparison.
A given inspection system is programmed to inspect the pattern of a wafer die or field of view, typically referred to as the inspected pattern, and then to compare it to the identical pattern of a second wafer die or field of view on the same wafer, serving as the reference pattern, in order to detect any pattern irregularity or difference which would indicate the possible presence of a wafer defect. A second comparison between the previously designated inspected pattern and the like pattern of a third wafer die or field of view is performed, in order to confirm the presence of a defect and to identify the wafer die or field of view containing the defect. In the second comparison, the first wafer die or field of view is considered as a reference.
Fabrication of semiconductor wafers is highly complex and very expensive, and the miniature integrated circuit patterns of semiconductor wafers are highly sensitive to process induced defects, foreign material particulates, and equipment malfunctions. Costs related to the presence of wafer defects are multiplied several fold when going from development stages to mass production stages. Therefore, the semiconductor industry critically depends on a very fast ramp-up of wafer yield at the initial phase of production, and then achieving and controlling a continuous high yield during volume production.
Critical dimensions of integrated circuits on wafers are continuously decreasing, approaching 0.1 micron. Therefore, advanced semiconductor wafers are vulnerable to smaller sized defects than are currently detected. Current methods of monitoring wafer yield involve optically inspecting, in-process, wafers for defects and establishing a feedback loop, with appropriate parametric process control, between the fabrication process and the manufactured wafers. To detect smaller sized defects, optical inspection systems realize increasingly higher resolution by means of scanning wafers using increasingly smaller pixel sizes. Scanning a given sized wafer using increasingly smaller pixel sizes causes a corresponding increase in per wafer inspection time, resulting in decreased wafer throughput, and decreased statistical sampling in terms of the number of inspected wafers. Conversely, attempting to increase wafer inspection throughput by using current optical system pixel sizes results in reducing the effectiveness, i.e., resolution, of wafer defect detection.
In addition to decreasing critical dimensions of wafers, the semiconductor industry is in the process of converting from manufacturing 8-inch wafers to 12-inch wafers. Larger, 12-inch wafers have more than twice the surface area of 8-inch wafers, and therefore, for a given inspection system, inspection time per 12-inch wafer is expected to be twice as long as that per 8-inch wafer. Fabricating 12-inch wafers is significantly more expensive than fabricating 8-inch wafers. In particular, costs of raw materials of 12-inch wafers are higher than those of 8-inch wafers. One result of wafer size conversion, is that cost effective productivity of future wafer manufacturing will depend critically upon increasing speed and throughput of wafer inspection systems.
Automated wafer inspection systems are used for quality control and quality assurance of wafer fabrication processes, equipment, and products. Such systems are used for monitoring purposes and are not directly involved in the fabrication process. As for any principal component of an overall manufacturing system, it is important that a wafer inspection method and system of implementation be cost effective relative to the overall costs of manufacturing semiconductor wafers.
There is thus a need to inspect semiconductor wafers for wafer die defects, for wafers featuring larger sizes and smaller critical dimensions, at higher throughput than is currently available, and in a cost effective manner.
Automated optical wafer inspection systems were introduced in the 1980's when advances in electro-optics, computer platforms with associated software and image processing made possible the changeover from manual to automated wafer inspection. However, inspection speed, and consequently, wafer throughput of these systems became technology limited and didn't keep up with increasingly stringent production requirements, i.e., fabricating integrated circuit chips from wafers of increasing size and decreasing critical dimensions.
Current wafer inspection systems typically employ continuous illumination and create a two dimensional image of a wafer segment, by scanning the wafer in two dimensions. This is a relatively slow process, and as a result, quantity of on-line inspection data acquired during a manufacturing process is small, generating a relatively small statistical sample of inspected wafers, translating to relatively long times required to detect wafer fabrication problems. Slow systems of on-line defect detection result in considerable wafer scrap, low wafer production yields, and overall long turn-around-times for pin-pointing fabrication processing steps and/or equipment causing wafer defects.
A notable limitation of current methods and systems of wafer defect detection relates to registration of pixel positions in wafer images. Before wafer defects can be detected by standard techniques of comparing differences in pixel intensities of an image of a targeted or inspected wafer die to pixel intensities of an image of a reference wafer die, the pixel positions of the images of the inspected and reference wafer die need to be registered. Due to typical mechanical inaccuracies during movement of a wafer held on a translation stage, velocity of a wafer beneath a wafer inspection camera system is not constant. As a result of this, image pixel positions in the fields of a detector are distorted and may not be as initially programmed. Therefore, a best fit two-dimensional translation pixel registration correction is performed.
Prior art methods and systems of wafer defect detection, featuring a combination of continuous wafer illumination and acquiring a two dimensional image by either scanning a wafer in two dimensions using a laser flying spot scanner as taught in U.S. Pat. No. 5,699,447, issued to Alumot et al., or scanning a wafer in one dimension using a linear array of photo detectors as taught in U.S. Pat. No. 4,247,203, issued to Levy et al., requires a registration correction for all pixels or all pixel lines. These methods limit system speed, i.e. inspection throughput, and require substantial electronic hardware. Moreover, they result in residual misregistration, since no correction procedure is accurate for all pixels in an image. Residual misregistration significantly reduces system defect detection sensitivity.
An apparatus for photomask inspection is disclosed in U.S. Pat. Nos. 4,247,203, and 4,347,001, both issued to Levy at al. The apparatus described in those patents locates defects or faults in photomasks by simultaneously comparing patterns of adjacent dies on the photomask and locating differences. Using two different imaging channels, equivalent fields of view of each die are simultaneously imaged, and the images are electronically digitized by two linear diode array photo-detectors, each containing 512 pixels.
A two dimensional image of a selected field of view of each die is generated by mechanically moving the object under inspection in one direction, and electronically scanning the array elements in the orthogonal direction. During the detector exposure time, the photomask cannot be moved a distance of more than one pixel or the image becomes smeared. Therefore, the time to scan and inspect the photomask is very long. Since the photomask is moved continuously while the two dimensional images are generated, it is necessary that the photomask move without jitter and accelerations. This motion restriction requires a very massive and accurate air-bearing stage for holding and moving the photomask, which is costly. In addition, the wafer inspection apparatus of Levy et al. is capable of detecting 2.5 micron defects with 95% probability of detection on photomasks.
For critical dimensions of current semiconductor integrated circuits approaching 0.1 micron, this means that the inspecting pixel must be of similar size magnitude. Since inspection speed increases inversely with squared pixel size, the apparatus of Levy et al. would slow down by more than two orders of magnitude. Furthermore, it becomes impractical to implement a motion stage capable of meeting the required mechanical accuracies.
Wafer inspection has also been implemented using a single imaging and detection channel, based on a solid state camera using a two dimensional CCD matrix photo-detector, such as described in ‘Machine Vision and Applications’, (1998) 1: 205-221, by IBM scientists Byron E. Dom et al. A wafer inspection system designated as P300 is described for inspecting patterned wafers having a repetitive pattern of cells within each die, such as in semiconductor wafers for memory devices. The system captures an image field of view having 480 by 512 pixels.
The image processing algorithms assume a known horizontal cell periodicity, R, in the image, and analyze each pixel in the image by comparing it with two pixels, one pattern repetition period, R, away in either horizontal direction. Such a comparison of like cells within a single image is called a cell-to-cell comparison. The pixel under test is compared with periodic neighbors on both sides to resolve the ambiguity that would exist if it were compared with only a single pixel.
While this system is capable of simultaneously capturing a two dimensional image of the object under test, it is very slow in inspecting an entire wafer. Millions of image fields are needed to image an entire wafer. Since the system uses continuous illumination, such as is used with standard microscopes, the wafer must be moved, under the inspection camera, from field to field and stopped during the image exposure to avoid image smear. To reach another field, the mechanical motion stage carrying the wafer must accelerate and then decelerate to a stop at a new position. Each such motion takes a relatively long time and therefore inspecting a wafer typically takes many hours.
Increased illumination of the inspected area can be achieved using laser illumination. However, the nature of a laser beam, and especially its coherent nature, presents a number of problems when used as such an illuminating source in applications requiring a uniform illuminating flux over the inspected area, such as is required, for instance, in a wafer inspection system:    (i) Interference of light in the illumination optics creates non-uniformity in the illumination field.    (ii) Interference of the illuminated light by the structured pattern on the wafer creates artifacts in the image.    (iii) Surface roughness creates speckle that generates non-uniformity in the image.    (iv) The laser beam itself is generally not uniform. Using the laser beam directly as a light source creates non-uniform illumination.
In order to overcome items (i) to (iii) above, the effects of the coherent nature of the laser beam must be reduced and preferably eliminated completely. This process is known as coherence breaking.
There are two definitions related to the coherence of a laser beam:    (a) Spatial coherence, which is the phase relation between each spatial point in the laser beam spot. This allows different points in the spot to interact with each other in a destructive or constructive manner when the spot is illuminating a cyclic pattern or a rough surface. This quality depends mainly on the mode of the beam. For instance in the basic mode (TEM00) the spatial coherence is defined by the Gaussian profile of the beam.    (b) Temporal coherence, which is a measure of the time or the transit distance (the time multiplied by the speed of light in the medium concerned) over which the phase of the beam can be defined. This parameter depends on the type of laser and its spectral bandwidth. Thus, for instance, for the second harmonic of a Nd:YAG laser at 532 nm, the coherence length is about 8 mm in free space.
There are a number of methods described in the prior art for overcoming coherence effects in using laser illumination. Reference is made to the articles “Speckle Reduction” by T. S. McKecknie, pp. 123-170 in Topics in Applied Physics, Vol. 9, Laser Speckle and Related Phenomena, edited by J. C. Dainty, Springer Verlag (1984), “Speckle reduction in pulsed-laser photography” by D. Kohler et al., published in Optics Communications, Vol. 12, No. 1, pp. 24-28, (September 1974) and “Speckle reduction with virtual incoherent laser illumination using modified fiber array” by B. Dingel et al., published in Optik, Vol. 94, No. 3, pp. 132-136, (1993), and to U.S. Pat. No. 6,369,888 to A. Karpol et al., for “Method and Apparatus for Article Inspection including Speckle Reduction”.
The disclosures of all of the publications and documents mentioned in this section, and in other sections of this application are all herein incorporated by reference, each in its entirety.
The above-mentioned prior art solutions to the problem of coherence breaking variously have specific disadvantages, and it is an object of the present invention to attempt to overcome some of these advantages.